Voice Activity Detector (VAD) is a voice-triggered system wake-up mechanism. VAD enables the rest of the system to be in sleep mode in the absence of voice activity, thereby consuming very low power. VAD-based systems generate an interrupt on detection of voice activity. Figure 1-1 shows how the VAD responds to voice activity.

Figure 1-1 VAD Example

VAD is supported on all analog-to-digital converter (ADC) channels of the TLV320ADCx120/PCMx120-Q1 device family, that is, on analog and digital microphone channels. The digital microphone channel is preferred for low-power applications. This application note describes the operation of the VAD, the tunable parameters, and the device configurations required to support VAD.

The VAD block monitors the signal from a microphone channel for voice-like patterns and on detection of a matching pattern that triggers an interrupt. The VAD monitors for both an onset of voice-activity as well as the end of voice-activity. Both events can be mapped to interrupts.

The ADCx120 device also has the capability to automatically power-on and power-off based on the VAD interrupts. As an example, the ADCx120 system can be set up to monitor VAD activity on a digital microphone channel and then power on the analog microphone channels based on the VAD trigger.

There are two modes of VAD operation:

1. Auto Mode: The VAD triggers automatically to power on and power off the system.
2. User Mode: The VAD interrupts have to be monitored by the host and the system has to be powered on or off through I2C commands.

Note that in both Auto and User modes, the device also generates an interrupt on the GPIO or GPO pin which can be sent to an external DSP or SOC.

The salient features of VAD are as follows:

• No need for external clocks - The system reuses an internal oscillator tuned to generate the ADC and DSP clocks of the required frequency.
• Automatic switching between VAD mode and Record mode - The system switches from VAD mode to Record mode upon voice activity and switches back to VAD mode upon no voice activity in Auto VAD mode.
• Reconfigurable decision tree from RAM - This feature allows the default voice-activity decision tree to be updated with a different decision tree that has been trained for other acoustic event detections.

The VAD algorithm uses a decision tree classification-based algorithm for voice activity detection. The decision-tree parameters can be updated through coefficient writes so the VAD block can be reconfigured for other applications that need to use a decision-tree for detection. A 16-band non-uniformly-spaced IIR filter-bank is used for feature extraction. Feature selection parameters and the decision tree can be computed offline and updated through coefficient-memory writes. Figure 2-1 shows the signal processing chain for VAD.

Figure 2-1 VAD Block Diagram